Abstract. In this paper, the evaluation of random bit generators for security applications is discussed and the concept of stateless generator is introduced. It is shown how, for the proposed class of generators, the verification of a minimum entropy limit can be performed directly on the post-processed random numbers thus not requiring a good statistic quality for the noise source itself, provided that a sufficient compression is adopted in the post-processing unit. Assuming that the noise source is stateless, a straightforward entropy estimator to drive an adaptive compression algorithm is proposed. Examples of stateless sources are also discussed. Finally, an attack scenario against a noise source is defined and an effective approach to the attack detection is presented. The entropy estimator and the attack detection together guarantee the unpredictability of the generated random numbers.